Run Claude review automatically on every PR

[jlocke2]

Update: AGENTS.md is the canonical file (open standard, readable by Claude Code/Codex/Cursor); CLAUDE.md is a one-line import so Claude Code picks it up natively.

Adds a CLAUDE.md context file for AI agents (and new developers) as part of the team's agentic-readiness work: what this repo is, how it fits into Pipeline CRM, how to build/run/verify it, architecture, and the real pitfalls.

Everything in it was derived from this repo's code, scripts, and README — no invented commands. Please read it as reviewers who know this system best and correct anything that is inaccurate or stale.

🤖 Generated with Claude Code

Update: review runs pinned to Sonnet — cheaper per run, and the reviewing model differs from the model that authors most changes.

[sevgibson]

LGTM — uses the safe pull_request trigger (not pull_request_target, so fork PRs get no secrets), contents: read, and budget/turn caps.

[sevgibson]

Approved — flagging one fleet-wide hardening for the C rollout (this is one of the few public repos, so it matters here). The issue_comment trigger only checks contains(github.event.comment.body, '@claude') — no author_association gate, unlike claude-worker.yml. On a public repo a non-collaborator could trigger the secret-bearing run by commenting @claude (bounded by the $0.75/run cap, and the comment path checks out the base branch so there's no code-exec — it's a cost/abuse surface, not RCE). Suggest adding the same guard the worker uses to the comment branches:

contains(fromJSON('["OWNER", "MEMBER", "COLLABORATOR"]'), github.event.comment.author_association)

Only the public repos are exposed (mantle, pipeline-js-api-client, pipeline_api) — private repos are fine as-is, so this isn't blocking.